Abstract
Diffeomorphic image registration, offering smooth transformation and topology preservation, is required in many medical image analysis tasks. Traditional methods impose certain modeling constraints on the space of admissible transformations and use optimization to find the optimal transformation between two images. Specifying the right space of admissible transformations is challenging: the registration quality can be poor if the space is too restrictive, while the optimization can be hard to solve if the space is too general. Recent learning-based methods, utilizing deep neural networks to learn the transformation directly, achieve fast inference, but face challenges in accuracy due to the difficulties in capturing the small local deformations and generalization ability. Here we propose a new optimization-based method named DNVF (Diffeomorphic Image Registration with Neural Velocity Field) which utilizes deep neural network to model the space of admissible transformations. A multilayer perceptron (MLP) with sinusoidal activation function is used to represent the continuous velocity field and assigns a velocity vector to every point in space, providing the flexibility of modeling complex deformations as well as the convenience of optimization. Moreover, we propose a cascaded image registration framework (Cas-DNVF) by combining the benefits of both optimization and learning based methods, where a fully convolutional neural network (FCN) is trained to predict the initial deformation, followed by DNVF for further refinement. Experiments on two large-scale 3D MR brain scan datasets demonstrate that our proposed methods significantly outperform the state-of-the-art registration methods.
Introduction
Medical image registration—the process of aligning two images into a common coordinate system—is a cornerstone of clinical analysis. Ideally, these transformations should be diffeomorphic, meaning they are smooth, invertible, and preserve the topology of anatomical structures. Current solutions typically fall into two categories: traditional optimization methods (e.g., SyN, LDDMM) which are accurate but slow, and learning-based methods (e.g., VoxelMorph) which are fast but often struggle with fine-grained local deformations and generalization.
Methods
The core innovation of this work is the Neural Velocity Field (DNVF). Instead of storing a deformation as a discrete grid, the authors use a Multilayer Perceptron (MLP) to represent a continuous function.
Neural Velocity Field
The MLP takes a 3D coordinate $(x, y, z)$ as input and outputs a velocity vector $(v_x, v_y, v_z)$. To capture high-frequency anatomical details, the model utilizes sinusoidal activation functions, which avoid the spectral bias common in ReLU-based networks. To ensure the transformation is diffeomorphic, the velocity field is integrated over time using the Scaling and Squaring (SS) method, creating a smooth mapping.
Cascaded Framework (Cas-DNVF)
To accelerate convergence, a two-stage approach is proposed. A standard Fully Convolutional Network (FCN) first predicts a rough initial alignment. This "warm start" is then refined by the DNVF optimizer to capture pair-specific nuances.
Results
The method was evaluated on the OASIS (416 scans) and Mindboggle101 (101 scans) datasets.
Accuracy and Topology
DNVF and Cas-DNVF significantly outperformed state-of-the-art methods like SyN and SYMNet. On the Mindboggle dataset, Cas-DNVF achieved a Dice Similarity Coefficient (DSC) of 0.612, compared to 0.567 for the best baseline. Furthermore, the method maintained a near-zero ratio of negative Jacobian determinants, confirming successful topology preservation.
Efficiency
While slower than pure one-shot learning methods, DNVF is significantly faster than traditional optimization. It achieves superior results in approximately 48 to 72 seconds, providing a favorable trade-off between accuracy and computational cost.
Conclusion
DNVF bridges the gap between optimization-based and learning-based registration by using neural fields to model admissible transformations. While the current implementation faces some memory constraints due to the Scaling and Squaring step, the results demonstrate that per-pair optimization refined by a cascaded warm start offers a robust path toward high-accuracy medical image registration. Future work will explore end-to-end training and time-varying velocity fields.
Summary updated